Clarify supported and preferred devices in quantization docs

[cadamcat]

OpLinearQuantizerConfig describes granularity and block_size without linking the hardware guidance discussed in #2510. The compute-plan docstring also describes supported_compute_devices using the text for the preferred device.

Add a note beside granularity linking the compression guidance and MLComputePlan, and correct the supported-device description to match Apple's API documentation. The note distinguishes device eligibility, the plan's preferred device, and observed execution.

Validation: Sphinx 7.4.7 HTML builds completed before and after the patch on 181cbd341fa394ea53fd6895a2e0a46aaec27cb1, with example execution disabled. The same 86 baseline diagnostics remain, including existing table-parsing errors; no new diagnostics. Checked the two rendered sections and links, Python syntax and unchanged executable AST.